fix

orca-zhang · orca-zhang · commit ef0b5c43f117 · 2025-02-27T14:00:07.000+08:00
diff --git a/src/llama.cpp b/src/llama.cpp
@@ -614,9 +614,8 @@ static struct ggml_tensor * llm_build_kqv(
         ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
 
         if (n_embd_head_v < n_embd_head_k) {
-            cur = ggml_cont(ctx, ggml_view_3d(ctx, ggml_cont(ctx, cur), n_embd_head_v, n_head, n_tokens,
+            cur = ggml_cont(ctx, ggml_view_2d(ctx, ggml_cont(ctx, cur), n_embd_head_v*n_head, n_tokens,
                                ggml_element_size(cur) * n_embd_head_v_out,
-                               ggml_element_size(cur) * n_embd_head_v_out * n_head,
                                0));
         }
 

Original file line number	Diff line number	Diff line change
`@@ -614,9 +614,8 @@ static struct ggml_tensor * llm_build_kqv(`
`614`	`614`	`ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);`
`615`	`615`
`616`	`616`	`if (n_embd_head_v < n_embd_head_k) {`
`617`		`- cur = ggml_cont(ctx, ggml_view_3d(ctx, ggml_cont(ctx, cur), n_embd_head_v, n_head, n_tokens,`
	`617`	`+ cur = ggml_cont(ctx, ggml_view_2d(ctx, ggml_cont(ctx, cur), n_embd_head_v*n_head, n_tokens,`
`618`	`618`	`ggml_element_size(cur) * n_embd_head_v_out,`
`619`		`- ggml_element_size(cur) * n_embd_head_v_out * n_head,`
`620`	`619`	`0));`
`621`	`620`	`}`
`622`	`621`